Computer vision analysis of the body-pose similarity from two different subjects with the aim of the correct development of physical exercises.

Abstract

The aim of this project is to obtain a new Neural Network model capable to properly detect specific keypoints on human bodies. These keypoints will be later treated for real-time corrections in the fieldof sportsand rehabilitation exercises. Generally, keypoint detection models focus on unconstrained environments;training and testing images contain one or more people, they might be practicing different activities, people may not be centred in the image,different clothing and background,etc. However, this project has focused on a moreconstrained context. There is a specific activity that the main subject is practicing;sports. And, moreover,only one person appears in the middle of the imageand there are not object occlusions.In order to train the model, we have performed afine-tuning on an open-source model from PyTorchwith an open-source datasetthat focuses on sports;LSP. We have then analysedif by constraining the context, the neural network model performance isimproved.The conclusion that we have reached is that LSP dataset is not correlated enough with real case scenarios in which a person is practicing sports in front of acamera. The model we have trained is capable to estimate keypoints on LSP imageswith high accuracy but, despite of that, when the model is used ina real case scenario, model predictions have not been as good as expected.